Artificial neural network modelling ofmicrobial acclimation periods in soil

contaminated with petroleum hydrocarbonsA.M. SUCHaRSKI-TREMBLAY I and R. KOK2

JResearch & Technology Transfer Section, College d' Aljj'ed, Alji'ed, ON, Canada KOB lAO; and 2Department ofAgriculturaland Biosystems Engineering. McGill University, 21111 Lakeshore Road, Ste. Anne de Bellevue, QC. Canada H9X 3V9.Received 8 October 1996; accepted 14 March /997.

Suchorski-Tremblay, A.M. and Kok,. R. 1997. Artificial neuralnetwork modelling of microbial acclimation periods in soil con­taminated with petroleum hydrocarbons. Can. Agric. Eng.39:123-130. Artificial neural networks (ANNs) were used to modelthe acclimation of indigenous microorganisms in soil contaminatedwith diesel fuel or creosote. Acclimation data were obtained bymeasuring the appearance of 14C (as C02) mineralized from radiola­belled tracers that were added to soil microcosms. The ANNs weretrained and tested with the following inputs: incubation temperature,water content (as percentage of the soil's water holding capacity),addition or not of nitrogen and phosphorus fertilizers, and samplingtime. The ANN output was directly related to the cumulative percentof 14C recovered (%L I4C). The resultant ANN models were incorpo­rated into a user-friendly software package called AccliMat, writtenin QuickBASIC. With this package, the ANN models can be utilizedto calculate the amount of %L14C that will be recovered after a givennumber of days, or the number of days required to reach a given%L I4C.

Des reseaux de neurones artificiels (RNA) ont servi de modelesprevisionnels pour l'acclimatation de microorganismes indigenesdans un sol contamine par Ie diesel ou Ie creosote. Cette acclimata­tion a ete mesuree en fonction de la presence de l'isotope 14C (sousforme de C02), provenant de la degradation des traceurs radioactifsajoutes ades echantillons de sol. Les RNA ont ete soumis adifferen­tes conditions d'apprentissage suivant I'entree de cinq donnees: latemperature d'incubation, la teneur en eau du sol (par rapport a sacapacite de retention), l'ajout ou non d'engrais d'azote et de phos­phore, et Ie temps d'echantillonnage. Les valeurs produites par lesRNA etaient directement liees aux quantites de 14C recupcrces(%L I4

C) dans les microcosmes pectologique. Nos modcles RNA ontete incorpores aAccliMat, un progiciel relativement simple autiliser,ecrit en QuickBASIC. Dans Ie cadre de ce progiciel, les RNA peu­vent servir aprevoir Ie pourcentage de 14C rccupere dans un sol apresun certain nombre de jours, ou encore, le.temps requis pour atteindreun niveau donne de 14C (%L I4C).

INTRODUCTION

Successful biodegradation of hydrocarbon spills on soil isinfluenced by many factors: the level of contamination, thediversity and population of the indigenous microflora,weather conditions, soil nutrients, and time. When a spilloccurs, the microorganisms acclimate to the carbon sourceand begin to degrade it. During the acclimation, variousbiological processes take place in the soil organisms them­selves, as well as in the composition of the microbialcommunity. Riser-Roberts (1992) has classified these as:

I) induction or derepression of enzymes specific to thedegradation pathways of a particular compound,

2) increase in the number of organisms in the degradingpopulation,

3) random mutation in which new metabolic pathways areproduced that will allow degradation, and

4) adaptation of existing catabolic enzymes ... to the deg­radation of novel compounds.

Such processes are affected by environmental factors. Forinstance, enzyme induction and derepression may be tem­perature and nutrient dependent.

In this study, two situations were examined. In both, aradiolabelled tracer technique was used to follow the produc­tion of 14C (as C02) being mineralized from the tracers. Inthe first situation diesel fuel or creosote was added to a soilwhich had not been previously contaminated. In this caseboth the contaminant and the radiolabelled tracer were addedto the test soil and the microbial population had to adapt toutilize the hydrocarbon source provided. We refer to this asmicrobial acclimation. In the second situation hydrocarbonspillage onto the soil had occurred over many years and themicrobes were already fully acclimated. We were interestedin how long it takes after spring thaw before the microbesbegin to degrade the hydrocarbons again and this situation istherefore referred to as post-thaw re-acclimation. For thiscase only the radiolabelled tracer was added to the test soil.

Data on microbial acclimation and post-thaw re-acclima­tion were obtained by monitoring 14C production (as C02),under various treatments. The data were then modelled withArtificial Neural Networks (ANNs) and a program with auser-friendly interface (AccliMat) was written to facilitateaccess to these ANN models. With the program a user caninput the acclimation time so as to find the cumulative per­cent of 14C recovered (%l:14C ), or input the %l:14Crecovered to find the corresponding time required.

OBJECTIVES

The objectives of the project were: I) to generate a set ofmodels of the appearance over time of 14C which resultedfrom the microbial mineralization of 14C-octadecane when inthe presence of diesel fuel in soil, and from the mineralization

CANADIAN AGRICULTURAL ENGINEERING Vol. 39. No.2. April/May/June 1997 123

14 14 •of C-phenanthrene or C-pyrene when In the presence ofcreosote in soil and, 2) to make the models available tonon-experts via a user-friendly software interface.

BACKGROUND

Artificial neural networks

Artificial neural networks are structured similarly to biologi­cal brains so that they can mimic and approximate to someextent the "human" way of interpreting and solving prob­lems. An ANN's primary components are processingelements (PEs) which are joined together by connections in amanner analogous to biological neurons being joined withsynapses. The PEs are structures that receive a number ofinput signals which are attenuated by weights associated withthe connections they pass through. Within a PE, these attenu­ated inputs are added and the combined signal is fed into atransfer function. The output from the transfer function isthen fed through the connections to other PEs. Usually, forconvenience of description, the PEs and connections arearranged in layers. The base of the structure then consists ofthe input layer whose PEs receive the external inputs and,similarly, the top of the structure consists of the output layerwhose PEs emit the external outputs. In between are locateda number of hidden layers, each containing a number of PEs,as determined by the ANN designer. The PEs in the variouslayers are interconnected in a pattern that is controllable bythe ANN designer, although standard patterns are mostlyused (Kok et al. 1994; NeuralWare 1995a; Shukla et al.1996). For this study the fully-connected, feedforward pat­tern was used. This means that the layers are connected sothat the output from anyone PE is fed as input to all PEs inthe layer above it. The signal flow therefore proceeds linearlyfrom the input layer, through the hidden layer(s), and out viathe output layer.

In the operation of an ANN there are two phases: learning(or training) and recall (or testing). Much of the informationor knowledge stored in an ANN resides in the values of itsweights, the rest being implicitly represented by its structure.A network learns by adjusting these weights in response toinput/output data that are presented to it and various schemesmay be used to accomplish this. For example, ANNs can betrained using supervised backpropagation learning (Neural­Ware 1995a). In this, differences between the model's actualoutput and the ideal output it should have are used to correctthe weights according to a backpropagation scheme, using alearning rule, such as the Delta rule. After learning hasproceeded for a number of cycles and the network embodiesthe input-output relationship to a certain extent, the weightsmay be frozen and the network used for recall. In this phasethe ANN is presented with sets of inputs from which itgenerates output values. If ideal "source output" values, cor­responding to the recall inputs, are also available (e.g. fromthe learning data set) the ANN's performance can be judgedin terms of the degree to which the two sets of output valuesare in agreement.

ANNs can be used to model complex, nonlinear, and largedata sets. They can learn by example, without having a prioriknowledge of the subject being treated, and are robustenough to withstand noisy or incomplete inputs. ANNs have

124

been applied in many biosystems areas, such as the control offermentors and growth pattern recognition (NeuralWare1995b; Linko et al. 1995; Widrow et al. 1988). In this studyANN modelling was used because of the nature of the data,which strongly resemble bacterial growth curves.

Biological processes

Biological processes often provide a data-rich but knowl­edge-poor situation and the quality of automatic control thatis possible may be rather limited. In fermentation processesthere are three main factors that contribute to this: a lack ofreliable mathematical models to describe cell growth andmetabolite production, a lack of on-line sensors that candetect important process state variables, and the microorgan­isms themselves having a complex regulatory system withinthe cells, which the external control system can only affect bymanipulating the extracellular environment. Accordingly,limitations within the process and the availability of largequantities of data favor the utilization of ANNs in the controlsystem. They are, therefore, being increasingly exploited forsuch roles. For instance, Zhang et al. (1994) employed' theANN approach to model the production of Bacillusthurigiensis and then used that model to control the fermen­tation. Similarly, Linko et al. (1995) used an ANN model toestimate the consumption of sugar and the production oflysine by Brevibacterium f/avum in a fed-batch fermentationwith the intent to control amino acid production.

Kinetic model building

ANNs have also been used for kinetic model building. Forexample, Tyagiand Du (1992) developed an ANN model ofthe toxic effects of lead, chromium, zinc, nickel, and theirmixture on the growth of both acclimated and unacclimatedmicroorganisms. Their intent was to capture the kinetics ofheavy metal inhibition in an activated sludge process andthey chose the ANN approach because of the complex pat­terns involved, in terms of nonlinearity, multivariability, andinteraction between process variables. They also studied theeffects of nickel and chromium on acclimated and unaccli­mated sludge with neural models. As the kinetic patternsgrew in complexity, so did the neural models' architectures.For comparison, the data were also fed to a linear multipleregression method and the sum of squares of the errors of thetwo modelling approaches compared. The ANN's error wasone third that of the regression model.

MATERIALS AND METHODS

Soils and contaminants

This study involved the use of twelve soils in total. Thecompany cooperating on the project, Grace BioremediationTechnologies of Betz Dearborn, Inc. (Mississauga, ON) pro­vided us with 11 soils from different locations; the twelfthcame from nearby Alfred College (Alfred, ON). The soilswere divided into three groups of four, respectively being"clean", "historically contaminated" with diesel fuel, and"historically contaminated" with creosote (see Table I). Thesoils were named as follows. Clean: Guelph, Mississauga,Halton Clay, and Halton Sand; diesel contaminated: QuebecBunker C, Mississauga Refinery, Oakville Refinery, andCambridge Farm (the local soil); creosote contaminated:

SUCHORSKI-TREMBLAY and KOK

Table I: Elemenlal analysis ofelean and historicall)' contaminated (I'IC) soils calculated un a dry weight basis

Total C TOla11 Phosphorus POI~lssilll11 MagnesiumSoil Name (7c w/w) ('it w/w) (Illg P/kg) (mg K/kg) (nH! ,tg/kg)

elcHo Soils

Guelph 2.71 0.23 28 99.2 459

Miss;ssauga 2.92 0.16 7 150.5 384

Ha/101I Clay 1.99 0.02 9 61.3 232

Halton Sand 1.57 <0.01 6 64.8 36

He Soils: DieselQuebec: BI/llker C 7.07 0.1 (J 5 33.6 2310Mississol/ga Refinery 6.70 (J.06 3 20.6 17

Oakville Refinery 7,49 <0.01 I 3.3 44

Cambridge Farm 6.32 0.35 88 40.8 49

He Soils: Creosole

Alberta Clay 7.31 0.12 6 1'29.8 335New Brill/swick 6.04 0.13 7 89.9 64Quebec: Dumps;Ie 1.63 <0.01 2 14.6 8TrclIlo1l Indl/stria/ 8A4 <0.01 4 25.8 47

---

A/berra Clay. New BrulI.\wick. Quebec Dl/mpsile. (lnd Tn'lItonI"dllstrial. All were chosen based on site accessibility andpermission from the site owners to colleci samples. Fromc;'lch site a batch of about 40 kg was colleCled. All twelvebatches of soil were put through a #4 mesh sieve (4.75 111m

grid, USA Standard Tcsting Sieve) to remove large particles.stones and roots (Pramcr and Bartha 1972). The batches werethcn sub-sampled and this material (about 25 kg per soil) wasmixed one hundred times with a hand trowel. Thc prcparedclean soils were then stored in plastic bags. sCi.lIed and placedin a refrigerator at SoC. while the prepared historically con­taminated soils were stored frozen at _20ue 10 simulatewinlcr conditions. Also. for all prepared soils thc followingwere determined: the saturated water holding capacity(WHC) (Atlas and Bartha 1981): Ihe waler, N (Bremner andMulvaney 1982), P (Olsen and Sommers 1982), 10lal C (Nel­son and Sommers 1982). K (Knudsen et al. 1982) and Mg(Lanyon and Heald 1982) contents: total pctroleum hydrocar­bon (TPH) and polycyclic aromalic hydrocarbon (PAH)analysis was perfonned by Water Technology IntcrnationalCorp. (Burlington, ON). Thc latter two tests wcre performedto cnsure that the clean soils were clean and 10 dctcrminc to

what degree the historically COllli.lminatcd soils werc actu ..lIlypolluled.

To prepme the clean soils for Ihe acclimation experimcnls.separate samples of each. equivalent to 25 g dry mass. werecontaminated with 2S0 m2 dicsel fuel and 106 dpm (disinte­gri.llions per minute) 1~·C-octadecane. or with 150 mE.

6 14 -creosote and 10 dpm C-phenanlhrenc (radiolabels pur-chased from Sigma Chemical Co __ SI. Louis. MO). Toprcpare the hislorically contaminated soils only radiolabclledtraccrs were added. To the diesel-contaminated olles. 106

dpm I.JC-octadecane W.IS addcd pcr sample (cquivalenl to 25g dry mass). while to the creosolc-contaminated ones. 106

dpm 14e_pyrene was added. This resulted ill a lowl of sixteentypes of contaminated soil samples. Next. for both the accli~

malion and post-thaw re-acclimalion cxperimenls variousamounts of water. N. and P wcre added to these. as describedin the '"Microcosms" section below.

Microcosms

Soil acclimation was studied with microcosms. Each micro­cosm consistcd of a 250 mL biometcr llask having a centrallyfused vcnical glass cylinder (15 x 70 mm) into which a gastrap vial was placcd (s1~ Fig. I). The trap conlained S mL of2 N aOH 10 caplure C02 and was replaced when neces­sary (see below). Each llask contained a soil samplc prep'lrcdas described above (equivalent to 25 g dry mass of soil) towhich water and fcnilizer amcndmenls were added as re­quired. Enough water was added to bring the soil walercontcnt up 10 either 50% or 85% of \VHC: if N was added,

the Iota I sample I conlcnt was increased 10 1000 J.1g (withammonium nitratc): if P was added. the lotal sample P con-

CO, Trap Vialwith NaOH

Soil with14C hydrocarbonand contaminant

Fig. I. A soil microcosm.

CANADIAN AGRICULTURAL ENGINEERING Vol. 39. No.2. April/i\laylJullc 1997 125

tent was increased to 1000 Jlg (with triple superphosphatefertilizer). The flasks were then sealed with neoprene stop­pers and kept in the dark at either 10°C or 22°C to preventphotodegradation of the chemicals. Each microcosm wasduplicated, except for the negative controls (air dried soilswith radiolabel and/or hydrocarbon) making 528 microcosmsin total «(4 clean soils x 2 contaminants + 8 historicallycontaminated soils) x 2 temperatures x 2 WHC x 2 Namounts x 2 P amounts x 2 replicates) + 16 negative controlsnot used in modelling). As the radiolabelled tracer was min­eralized by the microbes, 14C02 was released and thenabsorbed by the gas trap. The trap vials were changed everyday or every two days for 22°C diesel microcosms and everyfive days for 22°C creosote and for both diesel and creosote10°C microcosms, for up to 150 days. The evaluation ofpetroleum hydrocarbon metabolism by means of radiola­belled tracer degradation was described and validated bySharabi and Bartha (1993).

The radioactivity of the trap contents was assessed withliquid scintillation counting (model LS6500; Beckman In­struments Inc., Fullerton, CA). The count sample wasprepared by mixing I mL of the trap solution, 1 mL of 2 NAcetic Acid, and 10 mL of "liquid scintillation cocktail"(Optiphase HiSafe II or Scintiverse, Fisher Scientific Inc.,Toronto, ON). The addition of the acid was required toneutralize the NaOH of the trap solution, so as to prevent thelatter from quenching the response of the fluors. The count­ing time was 10 min per diesel sample and 3 min per creosotesample. The data obtained consisted of 14C counts whichcorresponded to the amounts of tracer (and, hence, the quan­tity of contaminant) that had been metabolized during theperiod of residence of that particular vial of trap solution inthe microcosm. In this study acclimation and post-thaw re­acclimation were defined as being complete when at least10% of the tracer's radiolabel had been mineralized andtrapped (Briglia et al. 1994).

Data preparationFrom the 14C count data cumulative values were calculated,as a percentage of the total dose of radiolabelled tracer thathad been added to the microcosm (%~14C). These valueswere then further transformed to a log-inverse-fraction (Eq.1). This was done because better modelling accuracy wasrequired at lower fraction values, Le., we wanted the relativeerror of the model, rather than the absolute error, to stayconstant over the entire range of the fraction 14C recovered.

log - inverse - fraction = In ( 1O~ ) (1)

%L C

For creosote soils many of the %~14C values were zero(i.e., no 14C had been recovered at all), causing difficultieswith the transformation in Eq. 1. To facilitate the procedure,these values were adjusted to 0.001.

Artificial neural network modellingA number of different ANN architectures were tried to deter­mine which would best describe the data. As discussed in theBackground section, the ANNs used were of the fully-con­nected, feedforward type, but the way the inputs were fed to

126

the network, as well as the number of hidden layers (1, 2, or3), and the numbers of PEs in these were varied. Initially, theintent had been to create one or two ANNs, with each model­ling the combined data from a number of soils. In thatapproach the soil number was an input variable and we triedencoding it as a decimal quantity, as well as with severalbinary schemes. The quality of the resultant models was,however, insufficient to satisfy our needs. In the end, it wasdecided to model the soils entirely independently from oneanother, Le., to create a separate ANN for each soil type. Thisapproach also has the advantage that future models of newsoils can easily be added to the package. For all networks, theremaining inputs were two-value coded, except for time.Thus, for each of the sixteen ANNs the inputs were: incuba­tion temperature (0 = 10°C, 1=22°C), water content (0 = 50%of WHC, I = 85%), N amendment (1 = not added, 2 = added),P amendment (1 = not added, 2 = added), and time in hours.The output was the log-inverse-fraction of the 14C recovered,as defined by Eq. 1.

To train and test an ANN having a specific architecture,the procedure was to have the network learn from the sourcedata for 250,000 cycles and then to freeze its weights andhave it produce output values in recall mode using the sameinput values as had been used for the learning. The ANNrecall output values were then compared with the correspond­ing source output values. Subsequently, the ANN was furthertrained and tested at regular intervals, up to one millioncycles.

To compare two sets of outputs, average source valueswere first calculated for the duplicates (i.e., each microcosmhad been duplicated and there were two source output valuescorresponding to any given input set). The performance of anANN was then judged according to several measures thatwere calculated by comparing the recall and average sourceoutput values. The most important measure by which wejudged was the mean absolute difference between the recallvalue and the average source value. The arithmetic mean ofthe differences was calculated as well, to check for offset(under ideal circumstances this would be zero). The standarddeviation was used to evaluate the spread of the differencesand the root mean square value was calculated to emphasizelarge differences. The correlation coefficient between therecall and average source values was also computed.

After preliminary work, the most appropriate ANN wasfound to be the fully connected, feed-forward type with thehyperbolic tangent transfer function being used in all the PEsof its hidden layers. The ANNs were all trained with super­vised backpropagation learning using the Delta rule(NeuralWare 1995a). For diesel contaminated soils the bestarchitecture was found to be "5-11-11-1", denoting that therewere five PEs in the input layer, eleven PEs in each of twohidden layers, and one PE in the output layer (Fig. 2). Forcreosote contaminated soi Is, most final ANNs had the con­figuration 5-12-12-12-1 (i.e., three hidden layers eachcontaining twelve PEs) but for the Trenton Industrial soil5-13-13-13-1 was used and for the Guelph soil 5-14-14-14-1.

Software usedAll acclimation and post-thaw re-acclimation data were com­bined in a Lotus 1-2-3 spreadsheet (Lotus Development

SUCHORSKI-TREMBLAY and KOK

Fig. 2. Information now in a 5-11-1 I-I ANN.

number of days is then reported (0 the user.To validate the AccliMat software. two tests were run on

it. corresponding 10 the two types of request it can accommo­date. First. the program was used to generate recall output

values of log-invcfse-fraction of %lT4C recovered for allavailable soil. treatmen!. time combinations. These were thencompared to the corrcsponding values generated with theNcural\Varc package while the same trained ANN was resi­dent in its mcmory. Secondly. the iterationlllelhod was testedfor all soil-treatment combinations as follows: a) AccliMatwas used to lind three time values, by iteration. respectively.corresponding to one-quarter. one-half. and three-qu3rlcrs of

the difference bel ween maximum and minimum %~I.lC re­

covered: b) the %~14C recovered COlTcsponding to the threetime values was calculated with the ANN: c) these %~14Cvalues were then compared to the ones used to obtain the timevalues in step a) above. For example. for Mississ<luga soil,diesel acclimation, 22°C, 85% WHC, and no N nor P nmelld­mcn!. the difference between maximum and minimum

P TimeT H20 NInput Layer

"co2OUtput Layer

"""''lK<'''-.

Hidden 1Layers

~

I Start I -;

I I~

I Background information. I Bypass

-; I Collect inputs. I <-

-; Graph of chosen curve:A) What % 14C can be recovered after

(input) x days?B) How many days before (input) x% 14C is

recovered?

IA ! ! I B

Input number 01 days. Input % 14C.

Calculate % 14C Estimate the time to recoverusing appropriate % 14C by iteration.

ANN model.

Calculate % 14C usingappropriate model.

Calculate differencebetween % 14C target and% 14C predicted by model.

Difference < 0.001 No-;

! I Yes

Print results. Print results.

I Rerun or End. I -;

I Choice of input entry I I Hardcopy printed, I<- lor a rerun. Yes or No?

I End

Software written

Once the ANNs were suflicielHly trained.they were cxtracted from the NeuralWarepackage as C language subroutines. Thesewere then modified to BASIC format andincorporated ilHo the support software. Ac­c1iMal. which was written in QuickBASIC.Its nowchart is shown in Fig. 3. After theuser chooses a soil and treatments. Acclimatdisplays a graph of the modelled curvc. to~

gether with limits for the incubation timcand %~14C. At this poilH. the user is allowedto proceed in one of two ways. First. it ispossible to supply inputs corresponding to asoil and experimental treatment. togetherwith a timc. and request a %rl-lc value.Secondly. thc uscr c<ln specify treatmentconditions. soil. and a %~ 1olC. and request atime value. To accommodate the second re­quest, an iterative approach is used in theprogram. In the iteration mcthod the proce­dure starts with thc maximum time obtaincdfor that soil-tre'ltmen! combination and dc­creases this by rive dals every cycle.checking if the target %1: 1' C value has beenpassed. If it has been. the search direction isreversed (so that time will increase for suc­cessive iterations). the iteration interval setto one day and the time increased one day ata time until the target is passed again. The Fig. 3. Flowchart of AccliMat.

Corp., Cambridge. MA) and then exportedinto individual ASCII files that became thelearning and recall files for the AN mod­els. NeuralWarc's Professional II/Plus(Neural\Yare. Inc.. Pillsburgh. PAl was used10 create and train the ANI s. The measures10 compare Ihe recall output values 10 thesource values werc calculated with a pro­gram written in QuickBASIC 4.5.

CANADIAN AGRICULTURAL ENGINEERING Vol. 39. No.2. I\pril/MaylJIIllc IW7 127

%I.14C recovered is 13.6% so that the one-quarter, one-half,and three-quarters values are respectively 3.4, 6.7, and10.1 %. The AccliMat iteration procedure yielded 19, 34, and45 days as the time values correSjonding to these; with theANN model corresponding %I.I C values of 3.4, 6.8, and10.2% were obtained.

RESULTS AND DISCUSSION

Soils and contaminants

The results of the soil elemental analysis are summarized inTable I. The agricultural soils, Guelph and Cambridge Farm,had the highest initial amounts ofP and N; two of the histori­cally contaminated soils, Oakville Refinery (diesel) andQuebec Dumpsite (creosote), had the lowest available P, K,and Mg concentrations; the Quebec Bunker C soil had anoticeably high Mg concentration. Seven of the eight histori­cally contaminated soils had a total C content between 6%and 8.5%, whereas the Quebec Dumpsite soil had a low totalC content.

All eight historically contaminated soils had a high sandcontent (Quebec Bunker C - coarse sand, Mississauga Refin­ery - loamy fine sand, Oakville Refinery -sand, CambridgeFarm - very fine sand, Trenton Industrial -loamy sand, NewBrunswick - sandy loam, Alberta Clay -sandy loam, Quebec- sand), while three of the four clean soils were of a loamytype (Guelph - loam, Mississauga - clay loam, Halton Clay­loam) and one was sandy (Halton Sand - fine sand).

The results presented in Table II illustrate that, of the 8historically contaminated soils, Oakville Refinery and Mis­sissauga Refinery were the most contaminated withpetroleum hydrocarbons whereas Quebec Dumpsite was themost contaminated with polycyclic aromatic hydrocarbons.For all clean soils TPH values less than 100 J.lg/g soil wereobtained; only trace amounts of PAHs were found.

Table II: Total petroleum hydrocarbons (TPHs) andpolycyclic aromatic hydrocarbons (PAHs) inhistorically contaminated soils

Soil Name TPH (Ilg!g soil) PAH (Ilg!g soil)

DieselQuebec Bunker C 1220 traceMississauga Refinery 48,800 trace

Oakville Refinery 182,000 traceCambridge Farm 1110 trace

CreosoteAlberta Clay 852 11.0New Brunswick 174 21.2Quebec Dumpsite 2305 116.0Trenton Industrial 1224 20.9

Artificial neural networksResults of the ANN performance measure calculations arepresented in Table III. These are statistics of the log-inverse­fraction values (as defined in Eq. 1) and thereforedimensionless. The average mean absolute difference for all

128

soils was 0.1007. In terms of error in the %I.14C value, thiscorresponds to overestimation by a factor of 1.106 (e0.1007)or underestimation by a factor of 0.904 (e-0.1007). It must beremembered, however, that the ANNs were trained with thelog-inverse-fraction values, rather than the %I.14C values, sothat the relative error would remain fairly constant over theentire range. This is illustrated in Fig. 4 in which three sets ofHalton Sand source and ANN recall output values, corre­sponding to different treatments, are shown together. Theerror in the log-inverse-fraction remains fairly constant sothat the larger errors in the %I.14C would be obtained at thehigher values.

All the correlation coefficients were above 0.98 and thearithmetic mean differences were minimal, indicating thatthere was no systematic offset between the source and recalloutput values of the ANNs. The root mean square differenceswere similar in magnitude to the standard deviations of thedifferences, indicating that the differences themselves werenot very widely distributed. Also, the root mean square dif­ferences were similar in magnitude to the mean absolutedifferences, indicating that the distributions of the differ­ences were not skewed in any extreme manner. The bestmodel constructed was of Quebec Bunker C, and the worstwas of Guelph (respectively suffering from average overesti­mation factors of 1.052 and 1.182 in %I.14C).

Overall, the data from the diesel contaminated soils wereeasier to model than those of the creosote contaminated ones.In the ANNs the diesel soil data required only two hiddenlayers of 11 PEs each, while the creosote soil data requiredthree hidden layers of either 12, 13, or 14 PEs. Even so, theaverage mean absolute difference obtained for the diesel datawas 0.0942 whereas for the creosote data it was 0.1107. Thelarger error in the creosote models reflects the character ofthe data collected. Of the 256 creosote microcosms, 123never reached 1% I. 14C recovered within the 150 day maxi­mum allowed (Le., they remained far below our acclimationlimit of 10%). As well, for several creosote contaminatedsoils substantial 14C recovery occurred only under sometreatments, but not under others. Such differences occurredeven between replicate microcosms.

Software

In the first test of AccliMat it was used to find %I.14C for allexperimental combinations of soil, treatment, and time. Thelog-inverse-fraction values that resulted always differed lessthan 1.0x 10-5 from corresponding ones generated directlywith the NeuralWare package in recall mode. In the test of theiteration method the %I.14C recovered, corresponding to thetime values at one-quarter, one-half, and three-quarters of thedifference between maximum and minimum values wererespectively (average standard deviation): 0.3229 ± 0.4112,0.2587 ± 0.3277, and 0.1485 ± 0.1624.

CONCLUSIONS

The use of ANNs to model radiolabel tracer data was reason­ably successful; the models approximated the experimentaldata to a degree well within our expectations and needs.Subsequent to training, the extraction of the ANN modelfrom the NeuralWare package as a C subroutine was straight-

SUCHORSKI-TREMBLAY and KOK

Table III: Values of the ANN performance measures*

No. of Mean Root mean Arithmetic Standard Correlationdata absolute square mean of deviation of coefficient. r

Soil Name Contaminant lines difference difference differences differences

Clean SoilsGuelph Diesel 528 0.1461 0.1927 -0.0006 0.1929 0.9895

Guelph Creosote 456 0.1673 0.2365 -0.0013 0.2368 0.9982Mississauga Diesel 496 0.1048 0.1470 0.0014 0.1471 0.9938Mississauga Creosote 448 0.1247 0.2007 0.0015 0.2009 0.9986Halton Clay Diesel 796 0.1060 0.1722 0.0049 0.1722 0.9949Halton Clay Creosote 440 0.0763 0.1015 -0.0002 0.1016 0.9992

Halton Sand Diesel 760 0.0828 0.1147 0.0001 0.1148 0.9971

Halton Sand Creosote 440 0.0987 0.1545 -0.0013 0.1547 0.9977

HC SoilsQuebec Bunker C Diesel 336 0.0508 0.0695 0.0007 0.0696 0.9959Mississauga Refinery Diesel 246 0.0826 0.1051 0.0002 0.1053 0.9937Oakville Refinery Diesel 488 0.0791 0.1203 0.0000 0.1204 0.9971Cambridge Farm Diesel 224 0.0558 0.0715 0.0002 0.0717 0.9926Quebec Dumpsite Creosote 152 0.1017 0.1473 -0.0006 0.1478 0.9827Alberta Clay Creosote 208 0.0977 0.1435 -0.0002 0.1438 0.9983New Brunswick Creosote 200 0.0588 0.0745 0.0000 0.0747 0.9992Trenton Industrial Creosote 152 0.1296 0.1761 -0.0009 0.1767 0.9971

Totals for: Diesel 3874 0.0942 0.1350Creosote 2496 0.1107 0.1620All soils 6370 0.1007 0.1455

*Training results. The ANN was fitted separately for each soil-contaminant combination

8

lI---c----------..-.--.------- -----.-.--.-.------ --- --.-.-.-

forward, as was its translation intoBASIC. The AccliMat interface soft­ware facilitates access to the models,so that any user can easily obtain an­swers to both acclimation questions,Le., either time or %L I4C can be theunknown. The approach of modellingwith ANNs, extracting them, and hid­ing them behind a user-friendlyinterface is highly recommended tomake the results of experimental workavailable to consumers.

ACKNOWLEDGMENTS

The authors acknowledge the contri­butions of our partners in the project,Grace Bioremediation Technologiesand the University of Guelph, Depart­ment of Environmental Biology.Thanks is also accorded to Environ­ment Canada's Development andDemonstration of Site RemediationTechnology and Environmental Inno­vation Programs for funding thisproject.

32tire (thcUlulI'do d houre)

O'--------'----------L----__~ __J

o

...f 6 h-~---------·---·-----··-·----·····--·-···-----······-----.-.----

~ ",. " '"~ ~~~-:-~-&;-;-:-:~:-~--------------_._----_ .._.._---_._----.....-'tI A. It. ·"1.".

I' =~:~~~~~~=~=~~:=~=:=~=~~~~:=~~::~~~~~~~ A ~~~ Aa ' I •_ -a.. , .B 2 1-- - ------.3~:A.. -- -- ~-~-,.- - A"':li.. - ---- --- - - -

...~ b.&. A ....

f-o- -.- -----.-.•.••.----- -_ !L!..••..&.~-.&··~·~· TT:A""..i'.••I.-1i.-.I A.-,n:- .

C

Fig. 4. Source and ANN recall output values versus time for Halton sand forthree treatments: a) 10°C, 50% WHC, -N, -P; b) 10°C, 50% WHC, +N,+P; c) 22°C, 50% WHC, -N, +P.

CANADIAN AGRICULTURAL ENGINEERING Vol. 39. No.2. April/May/June 1997 129

REFERENCES

Atlas, R.M. and R. Bartha. 1981. Microbial Ecology:Fundamentals and Applications. London, England:Addison Wesley Publishing Co.

Bremner, I.M. and C.S. Mulvaney. 1982. Nitrogen-Total. InMethods of Soil Analysis, Part 2 -Chemical andMicrobiological Properties, 2nd ed, ed. A.L. Page,Chapter 31. Madison WI: American Society ofAgronomy, Inc., Soil Science Society of America.

Briglia, M., P.l.M. Middeldorp and M.S. Salkinoja-Salonen.1994. Mineralization performance of Rhodococcuschlorophenolicus strain PCP-l in contaminated soilsimulating on site conditions. Soil Biology andBiochemistry 26: 377-385.

Knudsen, D., G.A. Peterson and P.P. Pratt. 1982. Lithium,Sodium, and Potassium. In Methods ofSoil Analysis, Part2 - Chemical and Microbiological Properties, 2nd ed, ed.A.L. Page, Chapter 13. Madison WI: American Society ofAgronomy, Inc., Soil Science Society of America.

Kok, R., R. Lacroix, G. Clark, and E. Taillefer. 1994.Imitation of a procedural greenhouse model with anartificial neural network. Canadian AgriculturalEngineering 36: 117-126.

Lanyon, L.E. and W.R. Heald. 1982. Magnesium, Calcium,Strontium, and Barium. In Methods ofSoil Analysis, Part2 - Chemical and Microbiological Properties, 2nd ed, ed.A.L. Page, Chapter 14. Madison WI: American Society ofAgronomy, Inc., Soil Science Society of America.

Linko, S., T. Rajalahti and Y.-H. Zhu. 1995. Neural stateestimation and prediction in amino acid fermentation.Biotechnology Techniques 9: 607-612.

Nelson, D.W. and L.E. Sommers. 1982. Total carbon,organic carbon, and organic matter. In Methods of SoilAnalysis, Part 2 - Chemical and MicrobiologicalProperties, 2nd ed, ed. A.L. Page, Chapter 29. MadisonWI: American Society of Agronomy, Inc., Soil ScienceSociety of America.

130

NeuralWare. 1995a. Neural Computing, A TechnologyHandbook for Professional II/PLUS and New'alWorksExplorer. Pittsburgh, PA: NeuralWare, Inc.

NeuralWare. 1995b. Using New'aIWorks, A Tutorial forNeuralWorks Professional II/PLUS and NeuralWorksExplorer. Pittsburgh, PA: NeuralWare, Inc.

Olson, S.R. and L.E. Sommers. 1982. Phosphorus. InMethods of Soil Analysis, Part 2 - Chemical andMicrobiological Properties, 2nd ed, ed. A.L. Page,Chapter 24. Madison WI: American Society ofAgronomy, Inc., Soil Science Society of America.

Pramer, D. and R. Bartha. 1972. Preparation and processingof soil samples for biodegradation studies. EnvironmentalLetters 2: 217-224.

Riser-Roberts, E. 1992. Bioremediation of PetroleumContaminated Sites. Boca, FL: C.K. Smoley.

Sharabi, N.E. and R. Bartha. 1993. Testing of someassumptions about biodegradability in soil as measuredby carbon dioxide evolution. Applied and EnvironmentalMicrobiology 59: 1201-1205.

Shukla, M.B., R. Kok, S.O. Prasher, G. Clark and R. Lacroix.1996. Use of artificial neural networks in transientdrainage design. Canadian Agricultural Engineering 38:119-124.

Tyagi, R.D. and Y.G. Du. 1992. Kinetic model for the effectsof heavy metals on activated sludge process using neuralnetworks. Environmental Technology 13: 883-890.

Widrow, B., R.G. Winter, and R.A. Baxter. 1988. Layeredneural nets for pattern recognition. IEEE Transaction onAcoustics, Speech, and Signal Processing 36: 1109-1118.

Zhang, Q., J.F. Reid, I.B. Lirchfield, J. Ren and S.-W. Chang.1994. A prototype neural network supervised controlsystem for Bacillus thurigiensis fermentations.Biotechnology and Bioengineering 43: 483-489.

SUCHORSKI-TREMBLAY and KOK